Context-driven and risk aware machine learning framework for forecasting tunnel deformation in complex geological zones

Tunnels in complex geological zones and squeezing-prone ground often experience large deformations. To forecast tunnel squeezing deformation in such cases our data-centric method identifies the minimum viable deformation records per label, for stable and interpretable prediction. We propose a context-driven and risk aware machine learning framework for early-stage design having very limited number of deformation records. This would enable engineers to select appropriate support stiffness to maintain structural stability under given geological and site-specific conditions with contextual information. It ensures a meaningful contribution to model training with limited deformation records. The contextual information captures the probabilistic importance of geological and site-specific factors at both local and global scale. Compared to conventional methodologies, our approach demonstrates improved clarity and transparent decision-making. Adoption is recommended for better planning of support systems, promoting safer and more cost-effective tunnel designs even with limited site-specific deformation records. Ultimately, the study advances data-driven and risk-informed engineering practice for early-stage Tunnel design.